"""
Soft Actor-Critic with continuous action spaces (SAC)
Paper link: http://proceedings.mlr.press/v80/haarnoja18b/haarnoja18b.pdf
Implementation: TensorFlow2
"""
import numpy as np
from argparse import Namespace
from xuance.tensorflow import tf, tk, Module
from xuance.tensorflow.learners import Learner
[docs]
class AlphaLayer(Module):
def __init__(self):
super(AlphaLayer, self).__init__()
self.log_alpha = self.add_weight(name="log_of_alpha", shape=[], initializer=tf.zeros, trainable=True)
[docs]
class SAC_Learner(Learner):
def __init__(self,
config: Namespace,
policy: Module,
callback):
super(SAC_Learner, self).__init__(config, policy, callback)
if ("macOS" in self.os_name) and ("arm" in self.os_name): # For macOS with Apple's M-series chips.
if self.distributed_training:
with self.policy.mirrored_strategy.scope():
self.optimizer = {'actor': tk.optimizers.legacy.Adam(config.learning_rate_actor),
'critic': tk.optimizers.legacy.Adam(config.learning_rate_critic)}
else:
self.optimizer = {'actor': tk.optimizers.legacy.Adam(config.learning_rate_actor),
'critic': tk.optimizers.legacy.Adam(config.learning_rate_critic)}
else:
if self.distributed_training:
with self.policy.mirrored_strategy.scope():
self.optimizer = {'actor': tk.optimizers.Adam(config.learning_rate_actor),
'critic': tk.optimizers.Adam(config.learning_rate_critic)}
else:
self.optimizer = {'actor': tk.optimizers.Adam(config.learning_rate_actor),
'critic': tk.optimizers.Adam(config.learning_rate_critic)}
self.tau = config.tau
self.gamma = config.gamma
self.alpha = config.alpha
self.use_automatic_entropy_tuning = config.use_automatic_entropy_tuning
self.mse_loss = tk.losses.MeanSquaredError()
if self.use_automatic_entropy_tuning:
self.target_entropy = -np.prod(policy.action_space.shape).item()
if self.distributed_training:
with self.policy.mirrored_strategy.scope():
self.alpha_layer = AlphaLayer()
self.alpha = tf.exp(self.alpha_layer.log_alpha)
if ("macOS" in self.os_name) and ("arm" in self.os_name): # For macOS with Apple's M-series chips.
self.alpha_optimizer = tk.optimizers.legacy.Adam(config.learning_rate_actor)
else:
self.alpha_optimizer = tk.optimizers.Adam(config.learning_rate_actor)
else:
self.alpha_layer = AlphaLayer()
self.alpha = tf.exp(self.alpha_layer.log_alpha)
if ("macOS" in self.os_name) and ("arm" in self.os_name): # For macOS with Apple's M-series chips.
self.alpha_optimizer = tk.optimizers.legacy.Adam(config.learning_rate_actor)
else:
self.alpha_optimizer = tk.optimizers.Adam(config.learning_rate_actor)
@tf.function
def actor_forward_fn(self, obs_batch):
with tf.GradientTape() as tape:
_, actions_forward, log_pi = self.policy(obs_batch)
policy_q_1, policy_q_2 = self.policy.Qpolicy(obs_batch, actions_forward)
log_pi = tf.reshape(log_pi, [-1])
policy_q = tf.reshape(tf.math.minimum(policy_q_1, policy_q_2), [-1])
p_loss = tf.reduce_mean(self.alpha * log_pi - policy_q)
gradients = tape.gradient(p_loss, self.policy.actor_trainable_variables)
if self.use_grad_clip:
gradients, _ = tf.clip_by_global_norm(gradients, clip_norm=self.grad_clip_norm)
self.optimizer['actor'].apply_gradients(zip(gradients, self.policy.actor_trainable_variables))
else:
self.optimizer['actor'].apply_gradients(zip(gradients, self.policy.actor_trainable_variables))
return p_loss, log_pi, policy_q
@tf.function
def critic_forward_fn(self, obs_batch, act_batch, rew_batch, next_batch, ter_batch):
with tf.GradientTape() as tape:
action_q_1, action_q_2 = self.policy.Qpolicy(obs_batch, act_batch)
_, next_actions, log_pi_next = self.policy(next_batch)
target_q = self.policy.Qtarget(next_batch, next_actions)
target_q = tf.reshape(target_q, [-1])
log_pi_next = tf.reshape(log_pi_next, [-1])
target_value = target_q - self.alpha * log_pi_next
backup = rew_batch + (1 - ter_batch) * self.gamma * target_value
y_true = tf.stop_gradient(tf.reshape(backup, [-1]))
y_pred_1 = tf.reshape(action_q_1, [-1])
y_pred_2 = tf.reshape(action_q_2, [-1])
q_loss = self.mse_loss(y_true, y_pred_1) + self.mse_loss(y_true, y_pred_2)
gradients = tape.gradient(q_loss, self.policy.critic_trainable_variables)
if self.use_grad_clip:
gradients, _ = tf.clip_by_global_norm(gradients, clip_norm=self.grad_clip_norm)
self.optimizer['critic'].apply_gradients(zip(gradients, self.policy.critic_trainable_variables))
else:
self.optimizer['critic'].apply_gradients(zip(gradients, self.policy.critic_trainable_variables))
return q_loss
@tf.function
def alpha_forward_fn(self, log_pi):
with tf.GradientTape() as tape:
object_value = tf.stop_gradient(log_pi + self.target_entropy)
alpha_loss = -tf.math.reduce_mean(self.alpha_layer.log_alpha * object_value)
gradients = tape.gradient(alpha_loss, self.alpha_layer.trainable_variables)
self.alpha_optimizer.apply_gradients(zip(gradients, self.alpha_layer.trainable_variables))
return alpha_loss
@tf.function
def learn_actor(self, *inputs):
if self.distributed_training:
p_loss, log_pi, policy_q = self.policy.mirrored_strategy.run(self.actor_forward_fn, args=inputs)
return (self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, p_loss, axis=None),
self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, log_pi, axis=None),
self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, policy_q, axis=None))
else:
return self.actor_forward_fn(*inputs)
@tf.function
def learn_critic(self, *inputs):
if self.distributed_training:
q_loss = self.policy.mirrored_strategy.run(self.critic_forward_fn, args=inputs)
return self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, q_loss, axis=None)
else:
return self.critic_forward_fn(*inputs)
@tf.function
def learn_alpha(self, *inputs):
if self.distributed_training:
alpha_loss = self.policy.mirrored_strategy.run(self.alpha_forward_fn, args=inputs)
return self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, alpha_loss, axis=None)
else:
return self.alpha_forward_fn(*inputs)
[docs]
def update(self, **samples):
self.iterations += 1
obs_batch = tf.convert_to_tensor(samples["obs"], dtype=tf.float32)
act_batch = tf.convert_to_tensor(samples["actions"], dtype=tf.float32)
next_batch = tf.convert_to_tensor(samples["obs_next"], dtype=tf.float32)
rew_batch = tf.convert_to_tensor(samples['rewards'], dtype=tf.float32)
ter_batch = tf.convert_to_tensor(samples['terminals'], dtype=tf.float32)
info = self.callback.on_update_start(self.iterations,
policy=self.policy, obs=obs_batch, act=act_batch,
next_obs=next_batch, rew=rew_batch, termination=ter_batch)
q_loss = self.learn_critic(obs_batch, act_batch, rew_batch, next_batch, ter_batch)
p_loss, log_pi, policy_q = self.learn_actor(obs_batch)
if self.use_automatic_entropy_tuning:
alpha_loss = self.learn_alpha(log_pi)
alpha_loss = alpha_loss.numpy()
self.alpha = tf.math.exp(self.alpha_layer.log_alpha).numpy()
else:
alpha_loss = 0
self.policy.soft_update(self.tau)
info.update({
"Qloss": q_loss.numpy(),
"Ploss": p_loss.numpy(),
"Qvalue": tf.reduce_mean(policy_q).numpy(),
"alpha_loss": alpha_loss,
"alpha": self.alpha,
})
info.update(self.callback.on_update_end(self.iterations,
policy=self.policy, info=info,
log_pi=log_pi, policy_q=policy_q, p_loss=p_loss, q_loss=q_loss,
alpha_loss=alpha_loss, alpha=self.alpha))
return info